Feature Drift Research Articles

Rethinking deep clustering paradigms: Self-supervision is all you need

The recent advances in deep clustering have been made possible by significant progress in self-supervised and pseudo-supervised learning. However, the trade-off between self-supervision and pseudo-supervision can give rise to three primary issues. The joint training causes Feature Randomness and Feature Drift, whereas the independent training causes Feature Randomness and Feature Twist. In essence, using pseudo-labels generates random and unreliable features. The combination of pseudo-supervision and self-supervision drifts the reliable clustering-oriented features. Moreover, moving from self-supervision to pseudo-supervision can twist the curved latent manifolds. This paper addresses the limitations of existing deep clustering paradigms concerning Feature Randomness, Feature Drift, and Feature Twist. We propose a new paradigm with a new strategy that replaces pseudo-supervision with a second round of self-supervision training. The new strategy makes the transition between instance-level self-supervision and neighborhood-level self-supervision smoother and less abrupt. Moreover, it prevents the drifting effect that is caused by the strong competition between instance-level self-supervision and clustering-level pseudo-supervision. Moreover, the absence of the pseudo-supervision prevents the risk of generating random features. With this novel approach, our paper introduces a Rethinking of the Deep Clustering Paradigms, denoted by R-DC. Our model is specifically designed to address three primary challenges encountered in Deep Clustering: Feature Randomness, Feature Drift, and Feature Twist. Experimental results conducted on six datasets have shown that the two-level self-supervision training yields substantial improvements, as evidenced by the results of the clustering and ablation study. Furthermore, experimental comparisons with nine state-of-the-art clustering models have clearly shown that our strategy leads to a significant enhancement in performance.

Occurrence and Causes of Large dB/dt Events and AL Bays in the Pre‐Midnight and Dawn Sectors

AbstractA necessary condition for the generation of Geomagnetically Induced Currents (GICs) that can pose hazards for technological infrastructure is the occurrence of large, rapid changes in the magnetic field at the surface of the Earth. We investigate the causes of such events or “spikes” observed by SuperMAG at auroral latitudes, by comparing with the time‐series of different types of geomagnetic activity for the duration of 2010. Spikes are found to occur predominantly in the pre‐midnight and dawn sectors. We find that pre‐midnight spikes are associated with substorm onsets. Dawn sector spikes are not directly associated with substorms, but with auroral activity occurring within the westward electrojet region. Azimuthally‐spaced auroral features drift sunwards, producing Ps6 (10–20 min period) magnetic perturbations on the ground. The magnitude of is determined by the flow speed in the convection return flow region, which in turn is related to the strength of solar wind‐magnetospheric coupling. Pre‐midnight and dawn sector spikes can occur at the same time, as strong coupling favors both substorms and westward electrojet activity; however, the mechanisms that create them seem somewhat independent. The dawn auroral features share some characteristics with omega bands, but can also appear as north‐south aligned auroral streamers. We suggest that these two phenomena share a single underlying cause. The associated fluctuations in the westward electrojet produce quasi‐periodic negative excursions in the AL index, which can be mis‐identified as recurrent substorm intensifications.

A dimensionality reduction assisted evolutionary algorithm for high-dimensional expensive multi/many-objective optimization

Surrogate-assisted multi/many-objective evolutionary algorithms (SA-MOEAs) have shown significant progress in tackling expensive optimization problems. However, existing research primarily focuses on low-dimensional optimization problems. The main reason lies in the fact that some surrogate techniques used in SA-MOEAs, such as the Kriging model, are not applicable for exploring high-dimensional decision space. This paper introduces a surrogate-assisted multi-objective evolutionary algorithm with dimensionality reduction to address high-dimensional expensive optimization problems. The proposed algorithm includes two key insights. Firstly, we propose a dimensionality reduction framework containing three different feature extraction algorithms and a feature drift strategy to map the high-dimensional decision space into a low-dimensional decision space; this strategy helps to improve the robustness of surrogates. Secondly, we propose a sub-region search strategy to define a series of promising sub-regions in the high-dimensional decision space; this strategy helps to improve the exploration ability of the proposed SA-MOEA. Experimental results demonstrate the effectiveness of our proposed algorithm in comparison to several state-of-the-art algorithms.

Evolving cybersecurity frontiers: A comprehensive survey on concept drift and feature dynamics aware machine and deep learning in intrusion detection systems

Intrusion Detection Systems (IDS) have become pivotal in safeguarding information systems against evolving threats. Concurrently, Concept Drift presents a significant challenge in machine learning, affecting the adaptability and accuracy of predictive models in dynamic environments. Understanding the synergy between IDS and Concept Drift is crucial for developing robust security systems. The motivation behind this survey is driven by the emerging complexities in cyber threats and the dynamic nature of data streams, which necessitate advanced IDS capable of adapting to Concept and Feature Drift. Our analysis reveals a glaring omission in the existing literature—the integration of Concept Drift and Feature Drift within IDS. Most studies have focused on Concept Drift in a general context or on IDS but have yet to comprehensively consider the implications of data dynamics. This oversight has led to a fragmented understanding and suboptimal approaches to tackling modern cyber threats. To address this, we propose a comprehensive review that delves into the role of machine learning in IDS, explicitly focusing on Concept and Feature Drift. We have proposed a framework that includes all the necessary components for a drift-aware IDS. The framework incorporates dynamic feature selection, adaptive learning algorithms, and continuous monitoring techniques to handle Concept Drift and Feature Drift effectively. The survey highlights state-of-the-art methodologies and current challenges in integrating these concepts. The methodology involves an exhaustive analysis of published works from 2019 to 2024, comparing and contrasting various models and approaches. This includes a detailed examination of Concept Drift-aware IDS methods, dynamic feature selection techniques, and the impact of high dimensionality in IDS. These quantitative improvements underscore the necessity for developing adaptive and resilient IDS. The survey uncovers under-represented areas in current research, paving the way for future investigations. By highlighting these gaps and providing comparative data, the survey sets a clear direction for upcoming research efforts to foster the development of more dynamic and adaptable IDS solutions. The quantitative experimental evaluation of the proposed framework is planned to be conducted in a future article, where we will assess its effectiveness and performance in real-world scenarios.

Self-X heterogeneous attributed graph embedding-based product configuration framework for cognitive mass personalization

Cognitive mass personalization (CMP) is a promising manufacturing paradigm; equipped with cognitive capabilities like reasoning, CMP satisfies changeable needs via configuring personalized products at scale. In CMP, knowledge graphs (KGs) are exploited by smart product-service systems (SPSS) to support cognitive configuration/reconfiguration processes. However, the extant KG-enabled SPSSs are built upon fixed configurations and hybrid frameworks due to lacking a graph embedding (GE) model to render cognitive configuration decisions. In fact, GE is scarcely used in SPSS configuration, because it is not only compromised by the heterogeneity of KGs entailed by content-related specifications and complex structures but also influenced by the feature randomness and feature drift problems, which are triggered by accumulative errors and inconsistent objectives due to noisy assignments and different configuration tasks, separately. To address these limitations, a Self-X Heterogeneous Attributed Graph Embedding (SXHAGE) model is proposed in a Self-X architecture, which includes 1) self-attention graph attention networks, 2) a self-adaptive autoencoder, and 3) self-optimizing training objectives, to present heterogeneous data through jointly optimizing heterogeneous attributed entities and relations. A systematic SXHAGE-based configuration framework, in which product family design and configuration recommending are enabled by graph clustering and link prediction, is developed as a continuous updating loop to proactively configure personalized products. A real-world case study, i.e., configure personalized electric clippers via a web-based sustainable configuration platform, is performed to validate the applicability of the proposed framework in the CMP context. Moreover, extensive experiments on the case study dataset demonstrate the superiority of SXHAGE over the state-of-the-art algorithms, e.g., surpassing Deep Neighbor-Aware Embedding (DNENC) by 18 % in F1-score for graph clustering and by 5 % in ROC-AUC for link prediction.

Potential and Limitations of Anomaly Detection via Tower-Radar Monitoring of Wind Turbine Blades in Regular Operation with Convolutional Networks

The authors have already presented a first analysis of a field study on radar- based structural health monitoring of rotor blades on two operational wind turbines. This analysis will now be extended with measurements from a third operational wind turbine used as a case study for structural monitoring of individual rotor blades over different time intervals and blade modifications. More precisely, a sensor box with a 35 GHz FMCW radar and a highspeed camera was positioned at 90m height on the turbine tower, collecting radar data for structural information and videos for automatic ground truth labelling. After two months of collecting data, each individual rotor blade of that turbine was differently modified and measured again. The radar-facing surface of one rotor blade was partially coated with a microwave absorber foil, while an equivalent foil was applied to the backside of the second rotor blade. The third rotor blade was intentionally left unchanged for the sake of comparability. This study investigates potential limitations like general dataset complexity in radar-based anomaly detection for operational wind turbines. Furthermore, the discriminative ability of a convolutional neural network for classifying moving rotor blades is discussed when encountering feature drift through changing environmental and operational conditions, i.e. metadata like nacelle orientation, rotor blade pitch and rotation speed.

Evaluating the reliability of machine-learning-based predictions used in nuclear power plant instrumentation and control systems

The field of data-driven, neural-network-based machine learning (ML) has seen significant growth, with applications in various information and control systems. Despite promising real-world uses, the reliability of models remains questionable. Conventionally, reliability is assessed based on predictive fidelity, accuracy, and training effectiveness; however, quality developmental procedures and excellent training performance metrics do not guarantee operational reliability. Instead, an ML model's predictive performance depends on the training set's representativeness to the intended operational space. It is known that ML algorithms excel at interpolation but struggle with extrapolation tasks. Anomalies and feature drift can also reduce operational performance. Determining whether a new sample is an interpolation or extrapolation task involves out-of-distribution (OOD) detection for assessing its proximity to the existing training data. Thus, we present a real-time, model-agnostic individual prediction reliability evaluation method called Data Auditing for Reliability Evaluation (DARE) for applying OOD detection to the training dataset. We demonstrate on a feedforward neural network ML-integrated digital twin for predicting fuel centerline temperatures during loss-of-flow transients. DARE acts as a “data supervisor” in determining the model's applicability under different operating conditions. In this manner, we demonstrate how training data can serve as inductive evidence to support the reliability of ML predictions.

AVBAE-MODFR: A novel deep learning framework of embedding and feature selection on multi-omics data for pan-cancer classification

Integration analysis of cancer multi-omics data for pan-cancer classification has the potential for clinical applications in various aspects such as tumor diagnosis, analyzing clinically significant features, and providing precision medicine. In these applications, the embedding and feature selection on high-dimensional multi-omics data is clinically necessary. Recently, deep learning algorithms become the most promising cancer multi-omic integration analysis methods, due to the powerful capability of capturing nonlinear relationships. Developing effective deep learning architectures for cancer multi-omics embedding and feature selection remains a challenge for researchers in view of high dimensionality and heterogeneity. In this paper, we propose a novel two-phase deep learning model named AVBAE-MODFR for pan-cancer classification. AVBAE-MODFR achieves embedding by a multi2multi autoencoder based on the adversarial variational Bayes method and further performs feature selection utilizing a dual-net-based feature ranking method. AVBAE-MODFR utilizes AVBAE to pre-train the network parameters, which improves the classification performance and enhances feature ranking stability in MODFR. Firstly, AVBAE learns high-quality representation among multiple omics features for unsupervised pan-cancer classification. We design an efficient discriminator architecture to distinguish the latent distributions for updating forward variational parameters. Secondly, we propose MODFR to simultaneously evaluate multi-omics feature importance for feature selection by training a designed multi2one selector network, where the efficient evaluation approach based on the average gradient of random mask subsets can avoid bias caused by input feature drift. We conduct experiments on the TCGA pan-cancer dataset and compare it with four state-of-the-art methods for each phase. The results show the superiority of AVBAE-MODFR over SOTA methods.

A novel paradigm: Addressing real-time decomposition challenges in carbon price prediction

Carbon price prediction serves as a market indicator and economic driver, facilitating the adoption of more environmentally friendly production methods to achieve the emission reduction targets outlined in the Paris Agreement. In recent years, time series analysis and decomposition techniques have been widely applied to carbon price forecasting. However, few researchers have considered the issue of data feature drift caused by real-time decomposition. Specifically, as the sample size increases, data features undergo changes, rendering the trained models unable to fit the new data. This paper explains the underlying reasons for this phenomenon from new perspectives and proposes a novel paradigm that replaces the intrinsic mode function with a single-step fuzzy particle. This new paradigm corrects the issue of data feature drift and concentrates noise into a completely new sequence during the preprocessing stage. In the subsequent processing steps, the loss of correlation between the sequence and time lag, as well as the sequence and carbon prices, is addressed through multi-information association. This paradigm hybrid model can be applied to deterministic and interval multi-step predictions. Experimental results on the datasets from Guangdong and Hubei demonstrate that the proposed model outperforms other comparative models, achieving better predictive results than existing decomposition-based forecasting models. In the case of Guangdong, the normalized root mean square error (NRMSE%) for one-step, three-step, and five-step deterministic predictions are 2.18%, 2.51%, and 2.89%, respectively. The average interval score (AIS) for interval predictions are −0.357 and − 0.2375, respectively.

TiSPHiMME: Time Series Profile Hidden Markov Ensemble in Resolving Item Location on Shelf Placement in Basket Analysis

A market basket in principle, mines/applies a variety of rules that tries to associate a set of items sold together and thus, generate sales transaction data in volumes daily. Its outcome is to provide users with adequate data against the issue of unnecessary item stock-up or inventory stock-out; Thus, averting un-needed demurrage, and provide clients with better decision and improved services. With such itemset (as a basket) seen to be time-bound, client-behavior over time does framework forecast, product that are commonly purchased vis-à-vis the itemset combination called a basket. We must also account for change in the features of a product vis-à-vis a corresponding change in shelve placement of the item – even as consumers change about their selected itemsets combination(s). Thus, our study explores a time-clustering algorithm that exploits (and mines) the Delta Mall (ShopRites) datasets to examine purchase behavior, preferences, and the frequency of itemset combination for each customer. For this model, we generated an average of 162-rules; And the results showed that previous basket items by random customers allow the selection purchase of items of similar value as best combined due to its shelf-placement using the concept of feature drift. Keywords; TiSPHiMME, Time Series, Hidden Markov Ensemble, Basket Analysis, Internet, Algorithms

Interpretable Siamese dual attention enhancement transfer compound diagnostic model for unbalanced samples

The intelligent transfer diagnosis model is used to address the issue of feature drift caused by the changing working conditions of rotating parts in engineering. However, few models can perform transfer diagnosis on multiple unbalanced samples of rotating parts simultaneously, and even fewer models can visually enhance the domain-invariant features, making them more interpretable. To address these issues, we propose a novel interpretable Siamese dual attention enhancement transfer compound diagnosis model for unbalanced samples. The model can diagnose multiple rotating parts simultaneously and consists of a channel feature attention enhancement (CFAE) network, a fragment feature attention enhancement (FFAE) network, and a Siamese feature fusion (SFF) network. The CFAE network enhances features of different convolutional channels, the FFAE network improves segment features in various frequency domains, and the SFF network extracts domain-invariant features of diverse rotating components under varying working conditions. The model is validated using bearing fault data collected under different loads and planetary gear fault data obtained at varying speeds. Its diagnostic accuracy remains above 96.4%, and the diagnostic variance is controlled within 1.0%. The model has good interpretability for imbalanced sample domain-invariant features, providing an effective tool for interpretable transfer diagnosis in this compound engineering situation.

A contrastive variational graph auto-encoder for node clustering

Variational Graph Auto-Encoders (VGAEs) have been widely used to solve the node clustering task. However, the state-of-the-art methods have numerous challenges. First, existing VGAEs do not account for the discrepancy between the inference and generative models after incorporating the clustering inductive bias. Second, current models are prone to degenerate solutions that make the latent codes match the prior independently of the input signal (i.e., Posterior Collapse). Third, existing VGAEs overlook the effect of the noisy clustering assignments (i.e., Feature Randomness) and the impact of the strong trade-off between clustering and reconstruction (i.e., Feature Drift). To address these problems, we formulate a variational lower bound in a contrastive setting. Our lower bound is a tighter approximation of the log-likelihood function than the corresponding Evidence Lower BOund (ELBO). Thanks to a newly identified term, our lower bound can escape Posterior Collapse and has more flexibility to account for the difference between the inference and generative models. Additionally, our solution has two mechanisms to control the trade-off between Feature Randomness and Feature Drift. Extensive experiments show that the proposed method achieves state-of-the-art clustering results on several datasets. We provide strong evidence that this improvement is attributed to four aspects: integrating contrastive learning and alleviating Feature Randomness, Feature Drift, and Posterior Collapse.

Dual Compensation Residual Networks for Class Imbalanced Learning.

Learning generalizable representation and classifier for class-imbalanced data is challenging for data-driven deep models. Most studies attempt to re-balance the data distribution, which is prone to overfitting on tail classes and underfitting on head classes. In this work, we propose Dual Compensation Residual Networks to better fit both tail and head classes. First, we propose dual Feature Compensation Module (FCM) and Logit Compensation Module (LCM) to alleviate the overfitting issue. The design of these two modules is based on the observation: an important factor causing overfitting is that there is severe feature drift between training and test data on tail classes. In details, the test features of a tail category tend to drift towards feature cloud of multiple similar head categories. So FCM estimates a multi-mode feature drift direction for each tail category and compensate for it. Furthermore, LCM translates the deterministic feature drift vector estimated by FCM along intra-class variations, so as to cover a larger effective compensation space, thereby better fitting the test features. Second, we propose a Residual Balanced Multi-Proxies Classifier (RBMC) to alleviate the under-fitting issue. Motivated by the observation that re-balancing strategy hinders the classifier from learning sufficient head knowledge and eventually causes underfitting, RBMC utilizes uniform learning with a residual path to facilitate classifier learning. Comprehensive experiments on Long-tailed and Class-Incremental benchmarks validate the efficacy of our method.

Stream learning under concept and feature drift: A literature survey

<p class="Abstract">Stream data learning is an emerging machine learning topic, and it has many challenges. One of its challenges is the dynamic behavior or changes in the environment which leads to drifts. Two types of drift occur, namely, concept drift and feature drift. This article provides a survey on stream data learning with focusing on the issues of feature drift and the methods developed for handling it. After presenting the fundamental concepts and definition in this field, it provides an overview of the various models and methods developed for detecting feature drift and maintaining the validity of the machine learning models when the drift occurs. Furthermore, the article provides the generators used for creating dataset with feature drift to provide benchmarking for approaches of detecting or handling feature drift. The article provides also taxonomy of feature selection methods in both static and dynamic environment. It concludes that reinforcement-based models are promising for this task, and it lists various open challenges and future works in this area.</p>

Rethinking Graph Auto-Encoder Models for Attributed Graph Clustering

Most recent graph clustering methods have resorted to Graph Auto-Encoders (GAEs) to perform joint clustering and embedding learning. However, two critical issues have been overlooked. First, the accumulative error, inflicted by learning from noisy clustering assignments, degrades the effectiveness of the clustering model. This problem is called Feature Randomness. Second, reconstructing the adjacency matrix sets the model to learn irrelevant similarities for the clustering task. This problem is called Feature Drift. Furthermore, the theoretical relation between the aforementioned problems has not yet been investigated. We study these issues from two aspects: (1) there is a trade-off between Feature Randomness and Feature Drift when clustering and reconstruction are performed at the same level, and (2) the problem of Feature Drift is more pronounced for GAE models, compared with vanilla auto-encoder models. Thus, we reformulate the GAE-based clustering methodology. Our solution is two-fold. First, we propose a sampling operator <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\Xi$</tex-math></inline-formula> that triggers a protection mechanism against Feature Randomness. Second, we propose an operator <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\Upsilon$</tex-math></inline-formula> that triggers a correction mechanism against Feature Drift by gradually transforming the reconstructed graph into a clustering-oriented one. As principal advantages, our solution grants a considerable improvement in clustering effectiveness and can be easily tailored to GAE models.

Visual transformer with stable prior and patch-level attention for single image dehazing

Single-image dehazing aims to recover blurred image details and improve image quality, which is a challenging ill-posed problem due to severe information degradation. In the image dehazing task, extracting local features from adjacent regions is particularly important. However, Transformer-based methods lack relative awareness of patch-level features. Furthermore, due to the sensitivity of self-attention to textcolorreddata distribution, the model suffers severe performance degradation when migrating from synthetic domain to real domain. To alleviate the above problems, we propose visual transformer with stable prior and patch-level attention (VSPPA) for image dehazing. Firstly, we propose a region-aware patch-level attention module to obtain the positional correlation between local patches and contexts, which can enhance the concentration of local patch-related features. Next, due to the instability problem caused by distribution shifts, we introduce dataset-independent prior to guide the transformer model, thereby preventing feature drift thus to improve the robustness of the model. Finally, domain-drift leads to insufficient dehazing when the model trained on synthetic data while migrates to the real environment, we come up with a introduce a patch filling strategy (PFS) for fuzzy data to narrow the domain gap and realize the generalization in real scenes. Extensive experiments show that the model achieves State-of-the-Art on the SOTS synthetic dataset and effective generalization to real-world scenarios.

Exploring the Interaction between Local and Global Latent Configurations for Clustering Single-Cell RNA-Seq: A Unified Perspective

The most recent approaches for clustering single-cell RNA-sequencing data rely on deep auto-encoders. However, three major challenges remain unaddressed. First, current models overlook the impact of the cumulative errors induced by the pseudo-supervised embedding clustering task (Feature Randomness). Second, existing methods neglect the effect of the strong competition between embedding clustering and reconstruction (Feature Drift). Third, the previous deep clustering models regularly fail to consider the topological information of the latent data, even though the local and global latent configurations can bring complementary views to the clustering task. To address these challenges, we propose a novel approach that explores the interaction between local and global latent configurations to progressively adjust the reconstruction and embedding clustering tasks. We elaborate a topological and probabilistic filter to mitigate Feature Randomness and a cell-cell graph structure and content correction mechanism to counteract Feature Drift. The Zero-Inflated Negative Binomial model is also integrated to capture the characteristics of gene expression profiles. We conduct detailed experiments on real-world datasets from multiple representative genome sequencing platforms. Our approach outperforms the state-of-the-art clustering methods in various evaluation metrics.

Metric Learning with Self-Adjusting Memory for Explaining Feature Drift

Metric Learning with Self-Adjusting Memory for Explaining Feature Drift

Interpretable SAM-kNN Regressor for Incremental Learning on High-Dimensional Data Streams

ABSTRACT In many real-world scenarios, data are provided as a potentially infinite stream of samples that are subject to changes in the underlying data distribution, a phenomenon often referred to as concept drift. A specific facet of concept drift is feature drift, where the relevance of a feature to the problem at hand changes over time. High-dimensionality of the data poses an additional challenge to learning algorithms operating in such environments. Common scenarios of this nature can for example be found in sensor-based maintenance operations of industrial machines or inside entire networks, such as power grids or water distribution systems. However, since most existing methods for incremental learning focus on classification tasks, efficient online learning for regression is still an underdeveloped area. In this work, we introduce an extension to the SAM-kNN Regressor that incorporates metric learning in order to improve the prediction quality on data streams, gain insights into the relevance of different input features and based on that, transform the input data into a lower dimension in order to improve computational complexity and suitability for high-dimensional data. We evaluate our proposed method on artificial data, to demonstrate its applicability in various scenarios. In addition to that, we apply the method to the real-world problem of water distribution network monitoring. Specifically, we demonstrate that sensor faults in the water distribution network can be detected by monitoring the feature relevances computed by our algorithm.

RRIFLoc: Radio Robust Image Fingerprint Indoor Localization Algorithm Based on Deep Residual Networks

Indoor localization is one of the most exciting research areas due to the increasing demand for location-based services (LBSs) in indoor environments. The fingerprint positioning method of the received signal strength indicator (RSSI) is widely used in indoor localization due to its simple deployment and low cost. However, since the RSSI is affected by indoor environment changes and the heterogeneous nature of devices, it is easy to cause fingerprint drift and disappearance of fingerprint features, resulting in low accuracy and weak robustness of indoor localization. In this article, we propose a robust indoor localization method that is calibrated-free of Wi-Fi image fingerprints, called the radio robust image fingerprint localization (RRIFLoc) algorithm. First, the signal strength difference (SSD) fingerprint and RSSI kurtosis are derived from the RSSI fingerprint. SSD and kurtosis alleviate the low positioning accuracy and weak anti-interference caused by fingerprint drift and the disappearance of fingerprint features. Second, the fusion of the RSSI, SSD, and kurtosis is constructed into a radio robust image fingerprint (RRIF). Finally, we build the RRIFLoc model using the generated RRIF and the deep residual network for location estimation. According to experiments on a public dataset, our method reduces the average location estimation error by 56.87% compared to state-of-the-art indoor fingerprint localization methods.